Using machine learning to train a shepherd dogŠtromajer, Tim (Avtor) Lebar Bajec, Iztok (Mentor) Demšar, Jure (Komentor) Ameri, Afshin (Komentor) collective behaviourshepherdingagent modelsartificial intelligencereinforcement learningsimulationDifferent organisms tend to form spontaneous and less predictable groups of individuals while doing everyday activities, such as eating or migrating. By understanding the rules of so called collective behaviour, we can learn how to control these groups to one's desires. Similar phenomena is of interest in many other domains like crowd control, cleaning the environment and other engineering problems. In this work, we focus on shepherding, which is an act of influencing or herding a flock of sheep using a shepherd dog. Here we create a model based on the existing shepherd dog models and then present some improvements to it to make it more realistic, such as limit the vision, add ability to hear other animals and implement a short term memory. We also adapt the model to allow multiple shepherd dogs to herd sheep at the same time. After that we present a model that is not based on some predefined rules, but is trained using reinforcement learning. Four different dog models are created, each able to observe the environment in a different way. The results show that the best model is the one that is using a ray casting method for observation.20222022-10-04 13:06:40Magistrsko delo/naloga141661VisID: 33881COBISS_ID: 124498691sl